OpenAI Accelerates Development of an Autonomous AI Researcher

OpenAI Redefines Its Ambitions with an Autonomous AI Researcher

OpenAI, the San Francisco-based technology company, has recently announced a major strategic shift in its research efforts. The company is now focusing on creating a fully automated AI researcher capable of tackling complex problems independently. This ambitious project is described as OpenAI's new "North Star," a central goal that will guide its research initiatives in the coming years. The company plans to integrate several research areas, including reasoning models, intelligent agents, and system interpretability.

An ambitious timeline has been set for this project. OpenAI aims to develop an "autonomous AI intern" by September, a system capable of autonomously solving a limited number of specific research problems. This intern will serve as a prototype for a fully automated multi-agent research system, which the company hopes to launch in 2028. The ultimate goal is to create an AI researcher capable of addressing problems that are too vast or complex for current human capabilities.

The potential application areas for this AI researcher are vast. They include mathematics and physics, where AI could formulate new proofs or conjectures, as well as life sciences like biology and chemistry. The tool could also be used to solve business and political dilemmas. In theory, any problem that can be expressed in text, code, or even scribbles on a whiteboard could be submitted to this tool, opening up an extremely wide range of applications.

Competition and Innovation

OpenAI has long been a leader in the field of artificial intelligence, particularly due to its advancements in large language models. These models are used daily by hundreds of millions of people. However, the company faces increasing competition from other major players like Anthropic and Google DeepMind. In this context, OpenAI's strategic decisions are crucial not only for its future but also for the overall evolution of AI.

The responsibility for defining these ambitious research goals largely rests on Jakub Pachocki, OpenAI's Chief Scientist. Pachocki has played a key role in the development of GPT-4, a groundbreaking language model launched in 2023, as well as in the reasoning models introduced in 2024. These technologies are now at the core of major chatbots and agent-based systems.

In a recent interview, Pachocki shared his vision for the future of OpenAI. He believes we are approaching an era where AI models can work consistently and continuously, much like humans. While human oversight remains essential for defining objectives, Pachocki envisions a future where a data center could host an entire research lab.

Towards Solving Complex Problems

The idea of solving the world's toughest problems through AI is not new. Industry leaders like Demis Hassabis from DeepMind and Dario Amodei from Anthropic have expressed similar ambitions. Sam Altman, OpenAI's CEO, has even mentioned goals as bold as curing cancer. However, Pachocki asserts that OpenAI is now better equipped to achieve these objectives thanks to recent advancements.

In January, OpenAI launched Codex, an agent application capable of generating code on the fly to accomplish various computing tasks. Codex can analyze documents, generate graphs, and even provide daily summaries of emails and social media. Similar tools, like Claude Code and Claude Cowork from Anthropic, have also emerged.

OpenAI indicates that the majority of its technical employees now use Codex in their daily work. Pachocki sees Codex as an embryonic version of the AI researcher that OpenAI aspires to create. The goal is to develop a system capable of functioning for longer periods with less human intervention. Pachocki explains that the idea is to design an automated research intern to whom one could delegate tasks requiring several days of human work.

A Promising Technological Evolution

The enthusiasm for creating systems capable of conducting prolonged scientific research is palpable. Doug Downey, a researcher at the Allen Institute for AI, emphasizes that the success of coding agents like Codex is a major source of inspiration. He wonders if similar systems could be developed for other scientific fields.

For Pachocki, the answer is affirmative. He believes we are already on the right track. The improvement of general capabilities in AI models naturally leads to systems that can work longer without assistance. He cites the evolution from GPT-3 in 2020 to GPT-4 in 2023 as an example of this progression. GPT-4 demonstrated the ability to work on problems for longer than its predecessor, even without specialized training.

Reasoning models have also provided a new impetus. By training LLMs to solve problems step by step and to backtrack in case of errors, their ability to work for extended periods has improved. Pachocki is confident that these models will continue to refine themselves.

OpenAI is also training its systems to work autonomously on complex tasks by providing them with specific examples, such as challenging puzzles from math and coding competitions. These exercises force the models to handle large volumes of text and to break problems down into multiple subtasks.

Challenges and Potential Risks

When asked about the risks associated with a system capable of solving large complex problems with little human supervision, Pachocki acknowledges that these questions are at the heart of discussions at OpenAI. He emphasizes that the acceleration of AI research, including research into AI itself, represents a major shift in the world.

Pachocki identifies several potential risks. The system could go off track, be hacked, or simply misinterpret its instructions. To mitigate these risks, OpenAI relies on thought chain monitoring, a technique where reasoning models note their actions in a virtual notebook. These notes allow researchers to verify that the model is behaving as intended.

In conclusion, while the development of an autonomous AI researcher by OpenAI promises to transform many sectors, it also raises important questions regarding safety and ethics. The ability of these systems to solve complex problems independently could have profound implications for the future of AI and society as a whole.